Research Presentation Session: Cardiac

RPS 2003 - Clinical applications of AI and ML in cardiovascular imaging

March 2, 14:00 - 15:30 CET

7 min
Deep learning denoising in cardiac CT imaging: improved image quality and workflow efficiency
Andreas Stefan Brendlin, Tübingen / Germany
Author Block: A. S. Brendlin, D. Wessling, J. Hofmann, J. Mück, B. Stenzl, S. Afat; Tübingen/DE
Purpose: To evaluate the impact of a deep learning denoising (DLD) algorithm on image quality, diagnostic confidence, and radiological workflows in the context of cardiac computed tomography angiography (CCTA).
Methods or Background: One-hundred patients (mean age 60±11 years) with a CCTA were included. Images were reconstructed using iterative reconstruction strength 2 (IR2) and the DLD algorithm. Place-consistent noise measurements were used to compare objective image quality. In addition, two blinded readers independently assessed subjective image quality, diagnostic confidence, sharpness, and contrast in a forced-choice setup. The results of these assessments were summarised for a semiquantitative overall quality score. Agatston-score and cardiac age were analysed semi-automatically for both datasets using proprietary software, before and after manually correcting the initial software output. The time required for manual corrections was measured for each reader to compare possible workflow benefits. Properly corrected mixed-effects analysis with post hoc subgroup tests was used. Spearman's correlation coefficient measured inter-reader agreement for the image quality analysis.
Results or Findings: Noise in IR2 was significantly higher than for DLD (22.00±2.32 versus 13.33±2.87 HU; p<0.001). DLD reconstructions had a significantly higher mean overall quality score than IR2 (3.5±1.0 versus 0.48±1.0, p<0.001) with a good inter-rater agreement (r≥0.790; p≤0.001). There were no significant differences between cardiac age results (p=0.517) and Agatson score values (p=0.486) of IR2 and DLD. However, the time required for manual corrections was significantly shorter for DLD than for IR2 (54±44 versus 35±31 seconds, p<0.001).
Conclusion: DLD significantly improves image quality in cardiac CT and substantially reduces the time required for manual corrections in cardiac age assessment.
Limitations: Identified limitations were: (1) the single-centre design; (2) the retrospective design; (3) the specific hardware and software setup.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The institutional review board approved retrospective eligibility analysis and data collection for this single-centre study's purpose with a waiver for the need for informed consent (reference number: 167/2022BO2).
7 min
Impact of AI-enabled motion compensation algorithm on coronary computed tomography angiography image quality
Giuseppe Stancanelli, Rome / Italy
Author Block: G. Stancanelli, L. Dominici, L. Conia, G. C. Pambianchi, C. Catalano, N. Galea; Rome/IT
Purpose: Motion artefacts remain a major limitation of coronary computed tomography angiography (CCTA), especially in patients with high cardiac frequency or rhythm variability and in coronary segments subject to motion, even in relatively quiescent phases of the cardiac cycle.
The aim of our study is to evaluate the impact on image quality of an AI-enabled interaction-free motion compensation reconstruction algorithm (MCR) compared to a standard filtered back projection reconstruction (FBP).
Methods or Background: Fifty patients underwent CCTA on a 128-slice scanner (Incisive CT, Philips) with an ECG modulated retrospective acquisition protocol. Raw datasets were reconstructed during the telediastolic phase of the cardiac cycle using a standard FBP algorithm and processed on an offline workstation to generate interaction-free MCR images (Precise Cardiac Suite, Philips).
The two image quality datasets were evaluated side by side by a reader with three years of experience blinded to the reconstruction technique. Image quality was graded per-segment and per-patient on a 1-4 scale based on the severity of the motion artefacts ("blurring", "winging" or "stairstep").
Results or Findings: Five hundred coronary artery segments were evaluated in both FBP and MCR reconstruction datasets; per segment coronary segmental image quality scores are reported in Fig 1. We observed a global statistically significant increase of mean scores after the application of the MCR algorithm (FBP: 2.75±1.04; MCR: 2.82±1.02 [p<0.01]. Overall, 15 out of 75 non-diagnostic segments were reclassified as diagnostic on MCR images, most of which (8/11) were on midRCA (p=0.02).
Conclusion: The application of MCR algorithm resulted in a global reduction of motion artefacts and an increase in image quality, with reclassification of non-diagnostic segments most evident on midRCA. This determined a better diagnostic performance of CCTA on segments most prone to motion artefacts.
Limitations: No limitations were identified.
Funding for this study: This study received no funding.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: No additional information provided by the submitter.
7 min
Automatic AI-based calcium scoring in cardiac and chest computed tomography: a validation study
Iris Hamelink, Zwolle / Netherlands
Author Block: I. Hamelink, Z. Nie, T. Kwee, M. Dorrius, P. M. A. Van Ooijen, R. Vliegenthart; Groningen/NL
Purpose: The extent of coronary artery calcium (CAC), assessed on computed tomography (CT), is a strong predictor of cardiovascular disease. The aim of this study is to validate the performance of an automatic AI system for quantifying CAC.
Methods or Background: 687 participants (59±4.8 years; 48.8% men) of the population-based ImaLife cohort were analysed for CAC. The Agatston score (AS) on cardiac and chest CT scans were quantified manually by a radiologist and automatically by an AI system (AI-Rad Companion Chest prototype, Siemens Healthineers). Agreement of manual and AI measurements was assessed by sensitivity and accuracy, Bland-Altman analysis and Cohen’s kappa for classification in AS strata (0; 1-99; 100-299; ≥300).
Results or Findings: Three participants were excluded due to incorrect manual measurement or a history of coronary stenting, resulting in 684 participants for evaluation. In cardiac CT, 200 (29%) participants showed no CAC when evaluated manually. 331 (48.4%) participants showed AS between 1 and 99, 92 (13.5%) participants between 100 and 299 and 61 (8.9%) participants ≥300. AI software showed a high sensitivity for CAC: 98.1% in cardiac CT (accuracy 97.2%) and 95.4% in chest CT (accuracy 92.1%). Bland-Altman analysis showed systematic bias of 2.3 and repeatability coefficient of 23.0 for AS on cardiac CT; and -0.3 and 38.0 for AS on chest CT. Cohen’s kappa for agreement in AS categorisation was 0.94 for cardiac CT and 0.87 for chest CT, with concordance in 96.0 and 91.4% of cases, respectively.
Conclusion: AI-based CAC scoring shows a high detection rate compared to manual evaluation, with excellent performance for risk classification. Performance is slightly better in cardiac CT than in chest CT.
Limitations: For both scan protocols, a low-dose, high-pitch scan protocol was used; it is unclear how generalisable results are for other scan protocols.
Funding for this study: The ImaLife study is supported by an institutional research grant from Siemens Healthineers and by the Ministry of Economic Affairs and Climate Policy by means of the PPP Allowance, made available by the Top Sector Life Sciences & Health to stimulate public-private partnerships.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the Medical Ethics Committee of the University Medical Center Groningen
7 min
AI-precision in cardiovascular risk assessment: non-gated chest CT coronary artery calcium scoring
Dan Mu, Nanjing / China
Author Block: D. Mu1, K. Yin1, W. Chen1, X. Chen2, B. Zhang1; 1Nanjing/CN, 2Shanghai/CN
Purpose: This study aimed to assess the performance of an artificial intelligence-based coronary artery calcium score (AI-CACS) algorithm on non-gated chest computed tomography (CT) images in differentiating risk categories for cardiovascular diseases.
Methods or Background: A prospective study enrolled 112 patients who underwent both chest CT and electrocardiogram (ECG)-gated non-contrast enhanced cardiac CT using the same equipment simultaneously. Chest CT images were reconstructed at three different thicknesses (1 mm, 3 mm, and 5 mm). The Agatston score, obtained semi-automatically from ECG-gated cardiac CT scans using a dedicated post-processing workstation, served as the reference. The AI-CACS software automatically derived the Agatston score from chest CT data. Correlations between AI-CACS and the reference Agatston score were calculated. The AI-CACS's performance in classifying risk categories, dichotomised at thresholds of 0, 100, and 400, was assessed.
Results or Findings: The AI-CACS showed strong correlations with the reference Agatston score for the three different slice thicknesses (1 mm: 0.973, 3 mm: 0.941, 5 mm: 0.834; all p < 0.001). Agreement in risk categories, assessed using kappa (κ) statistics, was substantial (κ = 0.868, p < 0.001), moderate (κ = 0.772, p < 0.001), and fair (κ = 0.412, p < 0.001) for 1 mm, 3 mm, and 5 mm slice thicknesses, respectively, with concordance rates of 91%, 84.8%, and 62.5%. When dichotomised at thresholds of 0, 100, and 400, the area under the curve for AI-CACS at the three slice thicknesses ranged from 0.785 to 0.996, 0.975 to 0.995, and 0.981 to 1.000, respectively.
Conclusion: The AI-CACS algorithm applied to chest CT images demonstrates promising performance in assessing cardiovascular disease risk. Using a 1 mm slice thickness for image reconstruction may yield the best results.
Limitations: A larger multi-centred, multi-vendor cohort study shall be conducted.
Funding for this study: No funding was obtained for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was ethically approved by the IRB number: 2022-547-01.
7 min
AI-based quantification of total and vessel-specific coronary artery calcifications in calcium scoring CT
Lilian Henriksson, Linköping / Sweden
Author Block: L. Henriksson, M. Sandstedt, A. Persson; Linköping/SE
Purpose: The objective of this study was to assess the correlation and agreement between fully automatic AI and standard semi-automatic evaluations by radiologists in calcium scoring CT (CSCT) examinations within the Swedish CardioPulmonary bio-image study (SCAPIS).
Methods or Background: A retrospective observational study involving 4567 CSCT exams carried out on a dual-source 128x128 slice CT scanner (Somatom Definition Flash, Siemens Healthineers, Forchheim, Germany) was conducted. Automatic AI measurements and semi-automatic measurements included Agatston score (AS), volume score (VS), mass score (MS), and number of lesions in the coronary arteries (LM+LAD, CX, RCA).
Results or Findings: Pearson correlation coefficients (r) for total AS, VS, and MS were r=0.989, 0.988, and 0.988 respectively. Intra-class correlation coefficients (ICCs) showed high levels of agreement: 0.994 for total AS, 0.998 for VS, 0.998 for MS, and 0.960 for the number of lesions. Bland-Altman analysis indicated minimal bias across all metrics. Weighted kappa for risk category classifications was 𝜅= 0.919, and the overall accuracy was 91.2%.
Conclusion: The study demonstrates excellent correlation and agreement between fully automated AI and radiologists' semi-automatic evaluations in CSCT examinations. The findings suggest that AI could be a reliable tool for calcium scoring, with implications for improving efficiency and standardisation in radiological assessments.
Limitations: Even though this is the largest study made so far regarding the accuracy of AI CSCT evaluations the study is limited by only including CSCT examinations performed on one type of CT scanner.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was ethically approved by the Swedish ethical review authority, regional ethical review board in Göteborg: DNR 2021-00441.
7 min
A deep learning algorithm for fully-automated myocardial infarct scar segmentation
Matthias Schwab, Innsbruck / Austria
Author Block: M. Schwab, M. Pamminger, C. Kremser, M. Haltmeier, A. Mayr; Innsbruck/AT
Purpose: Late gadolinium enhancement (LGE) cardiac magnetic resonance (CMR) imaging is considered the in vivo reference standard for assessing infarct size and microvascular obstruction (MVO) in ST-elevation myocardial infarction patients. As LGE distribution patterns can be quite complex and hard to delineate from the blood pool or epicardial fat, automatic segmentation of LGE CMR images is challenging. The aim of this work is to develop and evaluate a deep learning-based method that allows to perform LGE segmentation in a fully-automated way.
Methods or Background: A cascaded framework of two-dimensional and three-dimensional convolutional neural networks (CNNs), specialized on identifying myocardial scars, was trained on a training data set consisting of 224 patients. On a test data set including LGE CMR images from 152 examinations AI-based segmentations were compared to manual segmentations, which were performed according to the +5-SD method. Further, on a big data set of 1012 patients automatically calculated infarct volumes were correlated with maximum levels of creatine kinase and cardiac troponin obtained after acute myocardial infarction and successful primary percutaneous coronary interventions.
Results or Findings: Mean Dice coefficients between manual and CNN segmentations were 64.1% for LGE and 85.3% for MVO, respectively. Further, linear correlation between manually and automatically calculated infarct sizes was very strong (R=0.95, p<0.001). Good correlation between AI measured LGE volumes and biochemical measurements could also be found (creatine kinase: R=0.72, p<0.001; cardiac troponin: R=0.67 p<0.001).
Conclusion: Our fully-automated framework for LGE segmentation provides measurements that can compete with the very time-consuming manual segmentations.
Limitations: The limitations of the study are that for evaluation of the method, data from only one hospital (University Hospital Innsbruck) was used.
Funding for this study: Funding was provided by the Austrian Science Fund (FWF): DOC 110.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: The study utilised retrospective data analysis.
7 min
Artificial intelligence-based CT-derived FFR for the detection of haemodynamically significant coronary artery disease: a comparative study with dynamic stress CT myocardial perfusion imaging
Antoine Andary, Lille / France
Author Block: A. Andary, A. Rodriguez Musso, C. V. Gkizas, P. Carpentier, N. Abassebay, C. Lardemelle, B. Longere, F. Pontana; Lille/FR
Purpose: CT myocardial perfusion imaging (CT-MPI) combined with coronary CTA integrates coronary artery anatomy with inducible ischaemia at the cost of a higher radiation and contrast reinjection. The aim of this study was to evaluate the diagnostic performance of a deep-learning model of CT-derived fractional flow reserve (FFR-AI) for the detection of haemodynamically significant coronary artery disease (CAD) compared to CT-MPI.
Methods or Background: This retrospective study included 36 patients who underwent coronary CTA and dynamic stress CT-MPI on a third-generation dual-source CT system (SOMATOM Force, Siemens Healthineers). CT-MPI was performed when the maximal coronary stenosis was ≥50% (CAD-RADS≥3) or in the presence of stent, according to our centre’s practice after injection of regadenoson. Perfusion maps were interpreted by two radiologists by consensus. A perfusion defect was defined as a visually significant anomaly on the myocardial blood flow (MBF) map, in a coronary territory (windowing at 100 mL/100 mL/min by default). Curvilinear images of the main coronary arteries were then exported to CorEx model (version 1.0; Spimed-AI), which classified each of these arteries into two categories: FFR ≤0.8 or FFR >0.8.
Results or Findings: CT-MPI detected perfusion defects in 16 of 36 patients (44%). FFR-AI demonstrated a per-patient sensitivity, specificity, PPV, NPV and accuracy for the detection of hemodynamically significant stenosis of 100% (95% CI: 79%-100%), 50% (95% CI: 27%-73%), 61.5% (95% CI: 51%-71%), 100% and 72% (95% CI: 55%-86%), respectively. The areas under the ROC curve of FFR-AI were 0.75 (95% CI: 0.78-0.88).
Conclusion: FFR-AI provides high sensitivity and NPV for identifying haemodynamically significant CAD among patients with coronary stenosis ≥50%. FFR-AI could be used as a filter to avoid a subsequent CT-MPI and reduce radiation exposure and contrast reinjection.
Limitations: No limitations were identified.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study received institutional review board approval and written informed consent was obtained from all participants.
7 min
Machine learning explainable analysis for prediction of atrial fibrillation recurrence after catheter ablation using clinical and radiological variables
Alvaro Palazón Ruiz De Temiño, Alicante / Spain
Author Block: A. P. Ruiz De Temiño, J. M. Castro, M. J. Garfias, H. Trigueros Buil, A. Adarve Castro, D. F. Ferrández, B. Martínez-López; Alicante/ES
Purpose: Atrial fibrillation (AF) is a common arrhythmia with increasing prevalence and significant clinical impact. Catheter ablation has emerged as a treatment option for drug-resistant AF, with variable success rates. This study aimed to develop a machine learning-based model to predict AF recurrence after pulmonary vein ablation.
Methods or Background: A retrospective case-control study included patients who underwent first radiofrequency or cryoablation between 2017 and 2022. CT scans were used to measure left atrial volume (LAV), periatrial adipose tissue (PAT), interatrial adipose tissue (IAT), and (EAT) epicardial adipose tissue volume. Demographic, clinical, and recurrence data were collected. Feature selection and data preprocessing were conducted, followed by model training using three machine learning techniques. Model evaluation included accuracy, precision, recall, F1-score, and ROC/AUC. SHAP analysis was performed to interpret feature importance.
Results or Findings: Sixty nine patients were included. Recurrence occurred in 29% of patients. Persistent AF exhibited a higher risk of recurrence (OR 1.99, p<0.05). Radiological variables like left atrial, PAT and IAT volumes were significantly higher in recurrence cases. The logistic regression model including clinical and radiological variables (model A) achieved the highest average precision, accuracy, f1-score, and recall during cross-validation. Model A's accuracy in the testing group was 0.86, 0.66, and 0.86 and the AUC were 0.91, 0.87, and 0.92 using NN, NB, and LR respectively. SHAP analysis revealed varying feature importance across techniques in model A emphasizing the LAV, PAT and AF type.
Conclusion: This study presents two models incorporating adipose tissue measurements for predicting AF recurrence after pulmonary vein ablation with the potential of utilizing multimodal data in predicting post-ablation outcomes for AF patients.
Limitations: Sample size is limited, which might lead to overfitting. However, undersampling, scaling and cross-validation were employed as methods to mitigate this.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Patient data was handled confidentially and informed consent was waived by the ethics committee due to the retrospective nature and reliance on medical record review. Protocol registration number was PI2023-045.
7 min
Impact of super resolution deep learning reconstruction with 1024 matrix in inter and intra reader reproducibility of pre-TAVR CT measurements
Mickaël Ohana, Strasbourg / France
Author Block: A. Walch1, F. Tatsugami2, W. Fukumoto2, D. Touitou-Gottenberg1, A. Taniguchi3, K. Haioun3, K. Awai2, C. Roy1, M. Ohana1; 1Strasbourg/FR, 2Hiroshima/JP, 3Otawara/JP
Purpose: Reproducibility of aortic annulus sizing and aortic valve opening area planimetry on pre-TAVR cardiac CT is essential. Whether the use of a Super Resolution Deep Learning Reconstruction (SR-DLR) algorithm with increased matrix size could modify the inter and intra reader correlation of these measurements, particularly in case of heavily calcified aortic cusps, is unknown. Our primary objective is therefore to compare inter and intra reader reproducibility of aortic annulus and aortic valve area planimetry measurements between DLR and SR-DLR.
Methods or Background: Forty pre-TAVR CT with excellent image quality were retrospectively selected from two tertiary centers. Systolic phase was reconstructed with DLR in 512 and SR-DLR in 1024 matrixes. Four radiologists with different levels of expertise independently and randomly reviewed all 80 datasets to assess aortic annulus area and aortic valve planimetry. Two readers redid all measurements following a four week delay. Statistical analysis was performed using Bland-Altman plots and intraclass correlation coefficient (ICC).
Results or Findings: Interobserver agreement for aortic annulus area were excellent and similar between DLR (ICC 0.85, 95% CI 0.82-0.88) and SR-DLR (ICC 0.87, 95% CI 0.85-0.90). Interobserver agreement for aortic valve planimetry was higher with SR-DLR (ICC 0.90, 95% CI 0.86-0.92) than with DLR (ICC 0.83, 95% CI 0.80-0.85). This difference was more pronounced in the subgroup of patients with a heavily calcified aortic valve (calcium score >2000, n=24).
Intra-observer agreement for both measurements were slightly higher with SR-DLR.
Conclusion: SR-DLR with 1024 matrix could increase the reproducibility of aortic valve area planimetry, especially in heavily calcified aortic valve.
Limitations: Potential clinical implications of SR-DLR on device selection were not analyzed in this study.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was ethically approved by the IRB from Strasbourg University Hospital.
7 min
Enhancing preoperative risk assessment in noncardiac surgery (NCS): comparative evaluation of coronary CT angiography (CCTA), CT perfusion (CTP), and CT-derived fractional flow reserve (CT-FFR)
Federica Brilli, Rome / Italy
Author Block: F. Brilli, F. Catapano, C. Lisi, A. Caracciolo, M. Francone; Milan/IT
Purpose: Perioperative cardiovascular complications occur in approx. Three% of NCS hospitalisations. ESC guidelines recommended use of CCTA in low-to-intermediate likelihood of CAD, or in patients unsuitable for non-invasive functional testing undergoing non-urgent, intermediate/high-risk NCS. Our study sought to evaluate added value of CCTA-derived functional testing to predict revascularisation prior to NCS as compared to an anatomical-based strategy.
Methods or Background: Single-cohort prospective observational study including 55 symptomatic patients with stable angina who underwent CCTA prior to NCS; besides general CCTA and CTP contraindications, the presence of a planned invasive coronary angiography (ICA) for preoperative investigation before surgery was a major exclusion criterion. CT-FFR was performed using a ML-based algorithm for FFR simulation in all moderate to severe lesions. A ROC curve analysis was used to assess diagnostic performances of CCTA vs CTP vs CT-FFR in patients undergoing ICA after non-invasive testing.
Results or Findings: Significant stenoses were found in 20 participants and confirmed with ICA and FFR-ICA in moderate lesions. At ROC analysis, CTP had the largest AUC on a per-patient level (AUC = 0,84) compared with CT-FFR (0,41). The diagnostic accuracy of CTP and CT-FFR at patient-based analysis were 91% and 79%, respectively. The patient-based sensitivity, specificity, PPV, and NPV of CTP were 100%, 80%, 87% and 100%, whereas these values for CT-FFR (when using ≤0.80 as cutoff value) were 60%, 72%, 60%, and 88%. CCTA underperformed CTP for the diagnosis of flow-limiting coronary stenosis (accuracy at patient-based analysis: 77% vs 91%).
Conclusion: CTP offers a one-stop solution for assessing ischemic heart disease in NCS patients.
Limitations: Small sample size (55 patients) with stable angina, single-cohort observational design introducing potential selection bias and lacked long-term follow-up data. Larger cohort studies are needed to confirm CTP role in these patients.
Funding for this study: No funding was obtained for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Not applicable
7 min
Artificial intelligence for automated classification of coronary lesions from computed tomography coronary angiography scans (ALERT)
Victor Verpalen, Amsterdam / Netherlands
Author Block: V. Verpalen1, C. Coerkamp1, J. J. H. Henriques1, J-F. Paul2, N. R. Planken1; 1Amsterdam/NL, 2Paris/FR
Purpose: The aim of this study was to evaluate the diagnostic performance of a deep-learning model (DLM) for quantifying coronary stenosis on computed tomography coronary angiography (CTCA) using the Coronary Artery Disease-Reporting and Data System (CAD-RADS).
Methods or Background: This single centre retrospective study included 50 patients suspected of coronary artery disease (CAD). All CTCA examinations were obtained in routine clinical practice. Two expert readers and the DLM independently reassessed the CAD-RADS score per patient (n=50) and per vessel (n=150). Binary classification (CAD-RADS 0-2 or 3-5) and six group classification (CAD-RADS 0-5) were used for comparison among the human readers and between the readers and the DLM.
Results or Findings: Interhuman sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and Cohen’s kappa for detecting ≥50% stenosis (binary classification) were 86.4, 85.2, 82.6, 88.5, 85.7%, and 0.71 at patient level. Sensitivity, specificity, PPV, NPV, accuracy, and Cohen’s kappa of the DLM for detecting ≥50% stenosis were 100, 69.6, 75.0, 100, 84.1%, and 0.69 at the patient level for reader 1 and 100, 66.7, 71.4, 100, 81.8%, and 0.65 for reader 2 as reference, respectively. For the six group classification at patient level, interhuman agreement was 65.3% and weighted kappa 0.78. For the DLM vs reader 1 and reader 2 this agreement was 54.5 and 56.8%, the weighted kappa was 0.70 and 0.61, respectively.
Conclusion: Ruling out obstructive CAD (≥50% stenosis) by the DLM is safe, considering the 100% sensitivity. The DLM yielded promising results in CAD-RADS classification (0-5). This DLM has potential to support and alert CTCA-readers in clinical practice.
Limitations: The main limitation of the study is that the CAD-RADS distribution present in the study population does not necessarily reflect local clinical practice, which might influence the local performance of the DLM.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by Ethics committee Amsterdam UMC: 2023.0484,
7 min
Super-resolution deep learning reconstruction for improved image quality of myocardial late enhancement CT
Masafumi Takafuji, Tsu Mie / Japan
Author Block: M. Takafuji1, K. Kitagawa1, S. Mizutani2, A. Hamaguchi2, R. Kiso2, K. Sasaki2, Y. Funaki2, H. Sakuma1; 1Tsu Mie/JP, 2Matsusaka/JP
Purpose: Myocardial late enhancement CT (LE-CT) allows assessment of myocardial scar. Super-resolution deep learning image reconstruction (SR-DLR), which is trained on data acquired from ultra-high-resolution CT may improve image quality of LE-CT. The purpose of this study was to investigate image noise and image quality with SR-DLR compared with conventional DLR (C-DLR) and hybrid iterative reconstruction (hybrid-IR).
Methods or Background: We retrospectively analyzed 30 consecutive patients who underwent LE-CT using 320-row CT. The CT protocol consisted of stress dynamic perfusion CT, coronary CT angiography and LE-CT. All images were reconstructed using three different algorithms: SR-DLR, C-DLR, and hybrid-IR. Image noise, signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and overall image quality were compared. Overall image quality was assessed by five independent observers. Each observer had 30 points in each case and the points were allocated to LE-CT images reconstructed with the three different algorisms according to image quality. The scores were averaged across all observers.
Results or Findings: SR-DLR significantly decreased image noise by 33% compared to C-DLR (6.5±1.4 HU vs 9.7±1.7 HU, P<0.0001) and by 37% compared to hybrid IR (vs 10.4±2.8 HU, P<0.0001). SNR and CNR of LE-CT reconstructed using SR-DLR (SNR, 17.5±4.4; CNR, 4.6±0.8) were significantly higher than C-DLR (SNR, 11.4±2.8 p<0.0001; CNR, 3.1±0.6, p<0.0001) and hybrid-IR (SNR, 11.0±3.2, p<0.0001; CNR, 3.3±0.6, p<0.0001). SR-DLR significantly improved overall image quality of LE-CT compared to C-DLR (13.6±1.3 vs 8.6±0.7, p<0.0001) and hybrid-IR(vs 7.8±0.6, p<0.0001).
Conclusion: SR-DLR improved image noise, and image quality of myocardial LE-CT compared with C-DLR and hybrid-IR techniques. The SR-DLR approach has the potential to improve the assessment of myocardial scar by LE-CT and to lower the tube voltage and/or current of LE-CT, thus reducing the radiation dose of LE-CT.
Limitations: Not applicable.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the institutional review board in our institution, and written informed consent was obtained from each individual before enrolling in the study (reference number: 210604-5-2).

This session will not be streamed, nor will it be available on-demand!